From a8e475fdf4150baaec5d326582e1115af80cef26 Mon Sep 17 00:00:00 2001
From: indifferentketchup <samkintop@gmail.com>
Date: Sun, 7 Jun 2026 22:40:23 +0000
Subject: [PATCH] perf(llama): unshadow cache-type + spec-decoding flags for
 agent opt-in

KV cache quantization (--cache-type-k q4_0) and ngram speculative decoding
(--spec-type ngram-mod) are high-value llama.cpp features that improve VRAM
usage and tokens/sec. Removing them from the shadowing lists allows agents
to enable them via llama_extra_args.
---
 .../__tests__/llama-args-validator.test.ts    |  8 +++----
 .../inference/llama-args-validator.ts         | 23 +++++--------------
 2 files changed, 10 insertions(+), 21 deletions(-)

diff --git a/apps/server/src/services/__tests__/llama-args-validator.test.ts b/apps/server/src/services/__tests__/llama-args-validator.test.ts
index b1c2792..3794198 100644
--- a/apps/server/src/services/__tests__/llama-args-validator.test.ts
+++ b/apps/server/src/services/__tests__/llama-args-validator.test.ts
@@ -112,14 +112,14 @@ describe('stripShadowingFlags', () => {
     expect(result).toEqual(['-c', '4096']);
   });
 
-  it('strips cache flags by default', () => {
+  it('passes through cache flags (no longer shadowed)', () => {
     const result = stripShadowingFlags(['--cache-type-k', 'q8_0']);
-    expect(result).toEqual([]);
+    expect(result).toEqual(['--cache-type-k', 'q8_0']);
   });
 
-  it('strips spec flags by default', () => {
+  it('passes through spec flags (no longer shadowed)', () => {
     const result = stripShadowingFlags(['--spec-draft-n-max', '16']);
-    expect(result).toEqual([]);
+    expect(result).toEqual(['--spec-draft-n-max', '16']);
   });
 });
 
diff --git a/apps/server/src/services/inference/llama-args-validator.ts b/apps/server/src/services/inference/llama-args-validator.ts
index 78bd86f..127c408 100644
--- a/apps/server/src/services/inference/llama-args-validator.ts
+++ b/apps/server/src/services/inference/llama-args-validator.ts
@@ -131,23 +131,13 @@ export function isManagedFlag(flag: string): boolean {
 
 const SHADOW_CONTEXT = ['-c', '--ctx-size'];
 
-const SHADOW_CACHE = ['-ctk', '--cache-type-k', '-ctv', '--cache-type-v'];
+// Empty: agents should be able to opt into cache-type flags (lift analysis
+// found these are high-value features, not safety concerns).
+const SHADOW_CACHE: string[] = [];
 
-const SHADOW_SPEC = [
-  '--spec-default',
-  '--spec-type',
-  '--spec-ngram-size-n',
-  '--spec-ngram-size',
-  '--draft-min',
-  '--draft-max',
-  '--spec-draft-n-max',
-  '--spec-draft-n-min',
-  '--spec-draft-p-min',
-  '--spec-draft-p-split',
-  '--spec-ngram-mod-n-match',
-  '--spec-ngram-mod-n-min',
-  '--spec-ngram-mod-n-max',
-];
+// Empty: ngram speculative decoding is a performance feature agents should
+// be able to enable.
+const SHADOW_SPEC: string[] = [];
 
 const SHADOW_TEMPLATE = [
   '--chat-template',
@@ -160,7 +150,6 @@ const SHADOW_TEMPLATE = [
 // Shadowing flags that take no value — a boolean switch — so the stripper must
 // not also drop the following token.
 const VALUELESS_SHADOW_FLAGS: ReadonlySet<string> = new Set([
-  '--spec-default',
   '--jinja',
   '--no-jinja',
 ]);